
Codex Was Talked Into Backdooring Its Own Release Pipeline, Researchers Say
Ourtake
Nobody got backdoored this time, and the bounty hunters got paid, which is more than the agent would have done.
What happened
Security researchers at Depi reported a chain of flaws in OpenAI's Codex. A prompt injection in a GitHub issue workflow let an attacker run commands inside the Codex agent sandbox, which was then escaped to leak GitHub tokens. The tokens were used to poison CI/CD caches and plant a backdoored workflow. OpenAI accepted the report and fixed it, with a $1,000 reward.